What do you guess for Kraken vs M3?I could have guessed that. Of course 4 Zen5 + 8 Zen5C will beat 4P+4E in MT, especially considering that the latter doesn't have SMT either.
One could say SMT is cheating. It’s all tricks.AMD does it the hard way. A single core to serve 15W tablets to 500W serverside beasts while maintaining compatibility to an astonishingly large amounts of existing software.
Kraken is 8× Zen5? I'd guess Kraken is better than M3.What do you guess for Kraken vs M3?
Because the IOD use as much as 16W, so at 50W there s about 34W left for the cores while at 100W there s 84W left for the cores, do the maths, that s 2.47x less power for the cores at 50W compared to 100W TDP.Why is it that when I limit my 7950X to 50W it achieves a GB6 MT score that is less than half the score at 100W?
It's because there is SoC power vs core power. And while it's at 50% less total power, it goes from about 70W for the cores to play with to about only 20W.
On the low end of the Snapdragon Elite curve, don't you think it's possible there's a similar cause but to a lower degree?
I think @gdansk is being rhetorical.Because the IOD use as much as 16W, so at 50W there s about 34W left for the cores while at 100W there s 84W left for the cores, do the maths that s 2.47x less power for the cores at 50W compared to 100W TDP.
And that s at stock settings, if your RAM is ocked then the IOD will use 20W, that s 30W and 80W for the cores at 50W and 100W TDPs respectively, here the core power is 2.66x lower between the two points.
For the Snapdragon there can be such an effect but not of this scale because that s a monolithic and mobile dedicated chip where the uncore use a quite low power at idle.
There is no IO die for 7940HS, it’s a monolithic CPU. The only CPU with an IO die in that graph is MTL.I think @gdansk is being rhetorical. The graph above seems good for Snapdragon (and Intel) until you realise the AMD cpu has an IO die which makes the whole graph into a comparison of apples and oranges.
I edited my post for better precisions, i ll add that Qualcomm also pulled a completely flawed curve for AMD s GPU perf/watt wich is completely at odd with all tests done so far.I think @gdansk is being rhetorical. The graph above seems good for Snapdragon (and Intel) until you realise the AMD cpu has an IO die which makes the whole graph into a comparison of apples and oranges.
There is no IO die for 7940HS, it’s a monolithic CPU. The only CPU with an IO die in that graph is MTL.
Ahh my bad... read too fast. Cheers.There is no IO die for 7940HS, it’s a monolithic CPU. The only CPU with an IO die in that graph is MTL.
He mentioned the graph produced by Qualcomm, which happens to compare against a 7940HS.I suggest that you go consult an ophtalmologist, Gdansk used his 7950X as exemple, and that s what i am answering to.
Wide cores low clock, where have I heard that before ? hmmmI guess I'm totally wrong then.
He mentioned the graph produced by Qualcomm, which happens to compare against a 7940HS.
Yes, I don't remember the exact figures where this will happen but at least I can ballpark for my CPU. But my main point isn't about an IO die.I think @gdansk is being rhetorical. The graph above seems good for Snapdragon (and Intel) until you realise the AMD cpu has an IO die which makes the whole graph into a comparison of apples and oranges.
No clue, enlighten me please?Wide cores low clock, where have I heard that before ? hmmm
For all we know SDXE could be running at 600MHz at 5W and 3400MHz at 23W. So achieving 5.6x performance for 4.6x the power is within the realm of possibility when considering non-idle system performance (It's an MT benchmark after all).
I don’t get what this has to do with what my post. I wasn’t even addressing you.The MSI claw was tested at 28W against a 20W PHX and it s way behind in games where the CPU power is marginal in comparison of the GPU power, and yet MTL GPU use an efficient TSMC process, so you can imagine that at 30W CPU it is no match with its Intel 4 process against PHX2 TSMC s N4.
NBC measured roughly 16% higher perfs for PHX at about same 28W, this mean that MTL need rouhgly 38W to match PHX@28W, that s far from Qualcomm s showing the two chips being on par at 30W, actually they know that their most dangerous competitor technicaly wise is AMD, hence they tricked the numbers and showed the easily beatable Intel on a better light.
Ah of course. How could we forget!?And finally GB6 has some non-sensical results anyway. GB6 MT says a 8 core 5700X is within margin of error of a 12 core 5900 at 88W. Another reason why claiming Qualcomm is fabricating graphs defying the laws of physics because an AMD CPU has a poor showing is a bit of a reach.
Doubtful that it uses 1W at 600 MHz. At 600 MHz, that's like 15% of the peak performance. It's most likely in the range of <1W. You underestimate the power efficiency of the ARM cores.Just with those numbers i can tell you that it s a total fraud, even if the cores are dedicated a single watt at 600MHz
Same thing why cores aren't running 10GhzNo clue, enlighten me please?
Actually, better question: if widening the cores is just free efficiency at any frequency, why don't we have 10/12/14 wide decode?
Power consumption increases exponentially with increasing frequency.Heat?
Let me reiterate:Just with those numbers i can tell you that it s a total fraud, even if the cores are dedicated a single watt at 600MHz with the rest going to the uncore they would use at least 32W at 3400MHz, that s 5.66x the frequency, square this and you get the absolute best theorical case scenario that TSMC s process, or any other proces, cant yield at all, so that s even more than 32W.
Let me reiterate:
But of course it's a total fraud and not an overly reduced illustrative graph approximating reality
- You do not know the frequency - you can only guess.
- You do not know the core power - you can only guess.
- The scale isn't even labeled.
- GB6 MT doesn't correlate with transistor performance - 8 core 5700X and 12 core 5900 have same score, at same power.
- You're trying to do math with -no- concrete information by looking at a graph that doesn't correlate.